Voice agents have always sounded robotic, with that awkward pause while the app transcribes, thinks, then speaks. OpenAI just shipped three new realtime audio models in its API that close that gap, and they make building real voice agents genuinely practical instead of a science project.
What shipped
There are three pieces, now behind one unified Realtime API. GPT-Realtime-2 is the first voice model built on GPT-5-class reasoning, and instead of the old transcribe-then-respond pipeline it processes audio as a continuous stream. It can call several tools mid-conversation and even say “checking your calendar” out loud so the user knows it is working. GPT-Realtime-Translate does live translation across 70-plus input languages and 13 output languages, so two people can each speak their own language and hear the other in real time. GPT-Realtime-Whisper handles fast streaming transcription.
Why builders should care
The real shift is that reasoning, translation, and transcription now live behind one API with clear per-minute pricing. That deletes most of the glue code we used to write to stitch a voice agent together, and the latency is low enough to feel like a conversation, not a walkie-talkie.
If you have been putting off adding voice to your product, build a tiny prototype with GPT-Realtime-2 this week and feel the difference before you commit to an architecture.